mirror of
git://git.tartarus.org/simon/puzzles.git
synced 2025-04-21 08:01:30 -07:00
Files

introduces a few new utility functions in misc.c, one of which is the bitmap obfuscator from Mines (which has therefore been moved out of mines.c). [originally from svn r5992]
173 lines
5.1 KiB
C
173 lines
5.1 KiB
C
/*
|
|
* misc.c: Miscellaneous helpful functions.
|
|
*/
|
|
|
|
#include <assert.h>
|
|
#include <stdlib.h>
|
|
#include <string.h>
|
|
#include <stdio.h>
|
|
|
|
#include "puzzles.h"
|
|
|
|
void free_cfg(config_item *cfg)
|
|
{
|
|
config_item *i;
|
|
|
|
for (i = cfg; i->type != C_END; i++)
|
|
if (i->type == C_STRING)
|
|
sfree(i->sval);
|
|
sfree(cfg);
|
|
}
|
|
|
|
/*
|
|
* The Mines (among others) game descriptions contain the location of every
|
|
* mine, and can therefore be used to cheat.
|
|
*
|
|
* It would be pointless to attempt to _prevent_ this form of
|
|
* cheating by encrypting the description, since Mines is
|
|
* open-source so anyone can find out the encryption key. However,
|
|
* I think it is worth doing a bit of gentle obfuscation to prevent
|
|
* _accidental_ spoilers: if you happened to note that the game ID
|
|
* starts with an F, for example, you might be unable to put the
|
|
* knowledge of those mines out of your mind while playing. So,
|
|
* just as discussions of film endings are rot13ed to avoid
|
|
* spoiling it for people who don't want to be told, we apply a
|
|
* keyless, reversible, but visually completely obfuscatory masking
|
|
* function to the mine bitmap.
|
|
*/
|
|
void obfuscate_bitmap(unsigned char *bmp, int bits, int decode)
|
|
{
|
|
int bytes, firsthalf, secondhalf;
|
|
struct step {
|
|
unsigned char *seedstart;
|
|
int seedlen;
|
|
unsigned char *targetstart;
|
|
int targetlen;
|
|
} steps[2];
|
|
int i, j;
|
|
|
|
/*
|
|
* My obfuscation algorithm is similar in concept to the OAEP
|
|
* encoding used in some forms of RSA. Here's a specification
|
|
* of it:
|
|
*
|
|
* + We have a `masking function' which constructs a stream of
|
|
* pseudorandom bytes from a seed of some number of input
|
|
* bytes.
|
|
*
|
|
* + We pad out our input bit stream to a whole number of
|
|
* bytes by adding up to 7 zero bits on the end. (In fact
|
|
* the bitmap passed as input to this function will already
|
|
* have had this done in practice.)
|
|
*
|
|
* + We divide the _byte_ stream exactly in half, rounding the
|
|
* half-way position _down_. So an 81-bit input string, for
|
|
* example, rounds up to 88 bits or 11 bytes, and then
|
|
* dividing by two gives 5 bytes in the first half and 6 in
|
|
* the second half.
|
|
*
|
|
* + We generate a mask from the second half of the bytes, and
|
|
* XOR it over the first half.
|
|
*
|
|
* + We generate a mask from the (encoded) first half of the
|
|
* bytes, and XOR it over the second half. Any null bits at
|
|
* the end which were added as padding are cleared back to
|
|
* zero even if this operation would have made them nonzero.
|
|
*
|
|
* To de-obfuscate, the steps are precisely the same except
|
|
* that the final two are reversed.
|
|
*
|
|
* Finally, our masking function. Given an input seed string of
|
|
* bytes, the output mask consists of concatenating the SHA-1
|
|
* hashes of the seed string and successive decimal integers,
|
|
* starting from 0.
|
|
*/
|
|
|
|
bytes = (bits + 7) / 8;
|
|
firsthalf = bytes / 2;
|
|
secondhalf = bytes - firsthalf;
|
|
|
|
steps[decode ? 1 : 0].seedstart = bmp + firsthalf;
|
|
steps[decode ? 1 : 0].seedlen = secondhalf;
|
|
steps[decode ? 1 : 0].targetstart = bmp;
|
|
steps[decode ? 1 : 0].targetlen = firsthalf;
|
|
|
|
steps[decode ? 0 : 1].seedstart = bmp;
|
|
steps[decode ? 0 : 1].seedlen = firsthalf;
|
|
steps[decode ? 0 : 1].targetstart = bmp + firsthalf;
|
|
steps[decode ? 0 : 1].targetlen = secondhalf;
|
|
|
|
for (i = 0; i < 2; i++) {
|
|
SHA_State base, final;
|
|
unsigned char digest[20];
|
|
char numberbuf[80];
|
|
int digestpos = 20, counter = 0;
|
|
|
|
SHA_Init(&base);
|
|
SHA_Bytes(&base, steps[i].seedstart, steps[i].seedlen);
|
|
|
|
for (j = 0; j < steps[i].targetlen; j++) {
|
|
if (digestpos >= 20) {
|
|
sprintf(numberbuf, "%d", counter++);
|
|
final = base;
|
|
SHA_Bytes(&final, numberbuf, strlen(numberbuf));
|
|
SHA_Final(&final, digest);
|
|
digestpos = 0;
|
|
}
|
|
steps[i].targetstart[j] ^= digest[digestpos++];
|
|
}
|
|
|
|
/*
|
|
* Mask off the pad bits in the final byte after both steps.
|
|
*/
|
|
if (bits % 8)
|
|
bmp[bits / 8] &= 0xFF & (0xFF00 >> (bits % 8));
|
|
}
|
|
}
|
|
|
|
/* err, yeah, these two pretty much rely on unsigned char being 8 bits.
|
|
* Platforms where this is not the case probably have bigger problems
|
|
* than just making these two work, though... */
|
|
char *bin2hex(const unsigned char *in, int inlen)
|
|
{
|
|
char *ret = snewn(inlen*2 + 1, char), *p = ret;
|
|
int i;
|
|
|
|
for (i = 0; i < inlen*2; i++) {
|
|
int v = in[i/2];
|
|
if (i % 2 == 0) v >>= 4;
|
|
*p++ = "0123456789abcdef"[v & 0xF];
|
|
}
|
|
*p = '\0';
|
|
return ret;
|
|
}
|
|
|
|
unsigned char *hex2bin(const char *in, int outlen)
|
|
{
|
|
unsigned char *ret = snewn(outlen, unsigned char);
|
|
int i;
|
|
|
|
debug(("hex2bin: in '%s'", in));
|
|
|
|
memset(ret, 0, outlen*sizeof(unsigned char));
|
|
for (i = 0; i < outlen*2; i++) {
|
|
int c = in[i];
|
|
int v;
|
|
|
|
assert(c != 0);
|
|
if (c >= '0' && c <= '9')
|
|
v = c - '0';
|
|
else if (c >= 'a' && c <= 'f')
|
|
v = c - 'a' + 10;
|
|
else if (c >= 'A' && c <= 'F')
|
|
v = c - 'A' + 10;
|
|
else
|
|
v = 0;
|
|
|
|
ret[i / 2] |= v << (4 * (1 - (i % 2)));
|
|
}
|
|
return ret;
|
|
}
|
|
|
|
/* vim: set shiftwidth=4 tabstop=8: */
|